A0394
Title: Automated convolutional neural network and transformer for multi-class classification of three-dimensional brain images
Authors: Guan-Hua Huang - National Yang Ming Chiao Tung University (Taiwan) [presenting]
Abstract: Multiple model architectures that can effectively be applied to three-dimensional (3D) single photon emission computed tomography (SPECT) images are designed and developed, addressing the multi-class classification task of predicting stages of Parkinson's disease. The 3D SPECT imaging is treated as a sequence of two-dimensional (2D) image slice sequences, which are fed into a 2D convolutional neural network (CNN) or a transformer model to extract features. Then, these features were summarized in the second layer that implemented attention-mechanism-based models (including Transformer) to take into account the absolute or relative position information of input the slice. In the end, summarized features were used to derive a forecast for the disease stage. Finally, automated machine learning (AutoML) is applied to pre-processing selection and hyperparameter optimization tasks, which can quickly achieve the goal through automation and reduce the labor involved in manual selection. The analysis results showed that combining CNN for feature extraction and Transformer for summarization significantly improved the prediction accuracy and F1 scores of 3D image classification. Additionally, AutoML identified a combination of pre-processing approaches and hyperparameters that yielded a deep learning model with performance comparable to the best manually selected model, saving time and effort.
